Files
openteam/server/internal/proxy/passthrough.go
T
Sakurasan 79545aa48e fix: 兼容上游 SSE data: 后无空格格式,流式记账不再 0 消耗
火山方舟等上游流式返回 data:{...}(data: 后无空格),原解析只认
'data: '(带空格),导致 content 文本与 usage 全部漏解析:
- sseContentText 抽成 sseDataPayload,统一处理 data: 前缀
  (支持带/不带空格 + event:+data: 多行块)
- scanUsage 复用 sseDataPayload
- convert/stream.go parseLine 也支持 data:{ 无空格
- 新增回归测试覆盖无空格/带空格/多行块三种形态
2026-08-27 01:24:30 +08:00

653 lines
19 KiB
Go
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
package proxy
import (
"bufio"
"bytes"
"context"
"crypto/rand"
"encoding/hex"
"encoding/json"
"fmt"
"io"
"net/http"
"strconv"
"strings"
"time"
"github.com/gin-gonic/gin"
"github.com/openteam/server/internal/channel"
"github.com/openteam/server/internal/pkg/tokenizer"
"github.com/openteam/server/internal/store"
)
// newTraceID 生成请求 trace(用于日志与记账幂等 ref)。
func newTraceID() string {
b := make([]byte, 8)
_, _ = rand.Read(b)
return hex.EncodeToString(b)
}
// bodyReq 统一取出请求体并解析 model / stream 字段。
type bodyReq struct {
Model string `json:"model"`
Stream bool `json:"stream"`
}
// parseBody 读取并回填请求体,解析 model/stream。
func parseBody(c *gin.Context) (*bodyReq, []byte, error) {
body, err := io.ReadAll(c.Request.Body)
if err != nil {
return nil, nil, err
}
c.Request.Body = io.NopCloser(bytes.NewReader(body))
// 中断请求输入文本(token 估算用,仅存文本不立即分词)
c.Set("est_input_text", requestText(body))
br := &bodyReq{}
_ = json.Unmarshal(body, br) // 解析失败按空处理,直通仍可转发
return br, body, nil
}
// requestText 提取请求体中的用户输入文本(chat/messages 的 content、responses 的 input/instructions),
// 用于中断时估算输入 token。
func requestText(body []byte) string {
var m map[string]any
if json.Unmarshal(body, &m) != nil {
return ""
}
var parts []string
add := func(s string) {
if s = strings.TrimSpace(s); s != "" {
parts = append(parts, s)
}
}
if s, ok := m["instructions"].(string); ok {
add(s)
}
if s, ok := m["system"].(string); ok {
add(s)
}
switch input := m["input"].(type) {
case string:
add(input)
case []any:
for _, it := range input {
if im, ok := it.(map[string]any); ok {
if s, ok := im["content"].(string); ok {
add(s)
}
}
}
}
if msgs, ok := m["messages"].([]any); ok {
for _, msg := range msgs {
mm, ok := msg.(map[string]any)
if !ok {
continue
}
switch c := mm["content"].(type) {
case string:
add(c)
case []any:
for _, b := range c {
if bm, ok := b.(map[string]any); ok {
if s, ok := bm["text"].(string); ok {
add(s)
}
}
}
}
}
}
return strings.Join(parts, "\n")
}
// sseDataPayload 提取一条 SSE 的 JSON 载荷(去掉 data: 前缀与空白)。
// 兼容三种写法:
// - 单 data: 行:data: {...} 或 data:{...}(上游如火山方舟会省略 data: 后的空格)
// - event:+data: 多行块:转换器 eventLine 产出的块(event: xxx\ndata: {...} 拼在一个 []byte)
func sseDataPayload(line []byte) (string, bool) {
s := string(line)
idx := strings.LastIndex(s, "\ndata:")
if idx >= 0 {
s = s[idx+len("\ndata:"):] // 跳过 event: 头,落在 data: 之后
} else if strings.HasPrefix(s, "data:") {
s = strings.TrimPrefix(s, "data:")
} else {
return "", false
}
s = strings.TrimLeft(s, " ") // data: 后的可选空格
s = strings.TrimSpace(s)
if s == "" || s == "[DONE]" {
return "", false
}
return s, true
}
// sseContentText 提取一条 SSE 中的内容文本(chat delta.content / responses delta / messages delta.text)。
func sseContentText(line []byte) string {
s, ok := sseDataPayload(line)
if !ok {
return ""
}
var m map[string]any
if json.Unmarshal([]byte(s), &m) != nil {
return ""
}
// responses output_text.delta: {"delta":"..."}
if d, ok := m["delta"].(string); ok {
return d
}
// messages content_block_delta: {"delta":{"text":"..."}}
if dm, ok := m["delta"].(map[string]any); ok {
if t, ok := dm["text"].(string); ok {
return t
}
}
// chat chunk: {"choices":[{"delta":{"content":"..."}}]}(GLM 思考模型另有 reasoning_content)
if choices, ok := m["choices"].([]any); ok && len(choices) > 0 {
if c0, ok := choices[0].(map[string]any); ok {
if delta, ok := c0["delta"].(map[string]any); ok {
if t, ok := delta["content"].(string); ok {
return t
}
if t, ok := delta["reasoning_content"].(string); ok {
return t
}
}
}
}
return ""
}
// upstreamURL 组装上游地址:按协议选 base_url 再拼资源路径(见 store.Channel.UpstreamURL)。
func upstreamURL(ch *store.Channel, proto, path string) string {
return ch.UpstreamURL(proto, path)
}
// doProxy 通用代理(M5):遍历候选渠道,按需转换;可安全重试的失败自动故障转移。
func (g *Gateway) doProxy(c *gin.Context, cands []channel.Candidate, clientProto string, body []byte, stream bool, sink *usageSink) {
var lastStatus = http.StatusBadGateway
var lastBody = []byte("all upstream channels failed")
for _, cand := range cands {
ch := cand.Channel
plan, err := prepareUpstream(ch, clientProto, body, cand.UpstreamModel)
if err != nil {
lastStatus, lastBody = http.StatusInternalServerError, []byte("conversion error: "+err.Error())
continue
}
release, ok := g.ch.TryAcquire(ch)
if !ok {
continue // 渠道满载,溢出到下一个
}
written, retry, st, b := g.proxyOne(c, ch, plan, stream, sink)
release()
if written {
return
}
if !retry {
// 非重试性失败(如 400):透传上游错误体
c.Header("Content-Type", "application/json")
c.Data(st, "application/json", b)
g.recordError(c, ch, nil, now(), "upstream_http_"+strconv.Itoa(st))
return
}
lastStatus, lastBody = st, b
}
// 全部候选重试性失败
apiError(c, lastStatus, "upstream_error", string(lastBody))
g.recordError(c, nil, nil, now(), "all_channels_failed")
}
// proxyOne 对单个渠道执行一次代理。
// 返回:written=是否已写客户端响应;retry=是否可安全换渠道重试;status+respBody=失败信息。
func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan, stream bool, sink *usageSink) (written bool, retry bool, status int, respBody []byte) {
upKey, err := g.ch.UpstreamKey(ch)
if err != nil {
return false, true, http.StatusInternalServerError, []byte("failed to decrypt channel key")
}
upBody := plan.body
// 直通 chat 流式:注入 stream_options.include_usage,保证末块带 usage(OpenAI 行为)
if stream && plan.path == "/chat/completions" && plan.lineConv == nil && !bytes.Contains(upBody, []byte(`"include_usage"`)) {
var m map[string]any
if json.Unmarshal(upBody, &m) == nil {
m["stream_options"] = map[string]any{"include_usage": true}
if b, err := json.Marshal(m); err == nil {
upBody = b
}
}
}
ctx, cancel := context.WithTimeout(c.Request.Context(), time.Duration(ch.TimeoutMS)*time.Millisecond)
defer cancel()
req, err := http.NewRequestWithContext(ctx, http.MethodPost, upstreamURL(ch, plan.proto, plan.path), bytes.NewReader(upBody))
if err != nil {
return false, false, http.StatusInternalServerError, []byte("failed to build upstream request")
}
req.Header.Set("Content-Type", "application/json")
req.Header.Set("Authorization", "Bearer "+upKey)
req.Header.Set("Accept", c.GetHeader("Accept"))
if ua := c.GetHeader("User-Agent"); ua != "" {
req.Header.Set("User-Agent", ua)
}
if plan.path == "/messages" {
req.Header.Set("anthropic-version", "2023-06-01")
}
for _, h := range []string{"OpenAI-Organization", "OpenAI-Project", "OpenAI-Beta"} {
if v := c.GetHeader(h); v != "" {
req.Header.Set(h, v)
}
}
start := time.Now()
resp, err := g.hc.Do(req)
if err != nil {
if ctx.Err() == context.DeadlineExceeded {
return false, true, http.StatusGatewayTimeout, []byte("upstream request timed out")
}
return false, true, http.StatusBadGateway, []byte("upstream request failed: " + err.Error())
}
defer resp.Body.Close()
// 非 2xx:429/5xx 可重试;其余透传错误体
if resp.StatusCode < 200 || resp.StatusCode >= 300 {
errBody, _ := io.ReadAll(resp.Body)
if resp.StatusCode == http.StatusTooManyRequests || resp.StatusCode >= 500 {
return false, true, resp.StatusCode, errBody
}
return false, false, resp.StatusCode, errBody
}
c.Header("Content-Type", resp.Header.Get("Content-Type"))
c.Status(http.StatusOK)
if stream {
g.streamCopy(c, ch, resp.Body, start, plan.lineConv, sink)
} else {
g.copyAndCapture(c, ch, resp.Body, start, plan.bodyConv, sink)
}
return true, false, http.StatusOK, nil
}
// copyAndCapture 非流式:整体转发(可转换)+ 解析 usage + 记账。
func (g *Gateway) copyAndCapture(c *gin.Context, ch *store.Channel, r io.Reader, start time.Time, bodyConv func([]byte) ([]byte, error), sink *usageSink) {
data, err := io.ReadAll(r)
if err != nil {
apiError(c, http.StatusBadGateway, "upstream_error", "failed reading upstream response")
g.recordError(c, ch, nil, start, "read_error")
return
}
if usageRaw := extractUsage(data); usageRaw != nil && sink != nil {
sink.push(usageRaw)
}
out := data
if bodyConv != nil {
if converted, cerr := bodyConv(data); cerr == nil {
out = converted
}
}
_, _ = c.Writer.Write(out)
if _, ok := c.Get("raw_request"); ok {
c.Set("raw_response", string(data)) // 上游原始响应(未转换)
}
g.finishUsage(c, ch, start, store.UsageStatusSuccess, "")
}
// streamCopy 流式:边读上游 SSE 边写客户端,零缓冲转发;按 lineConv 转换;扫描 usage 记账。
func (g *Gateway) streamCopy(c *gin.Context, ch *store.Channel, r io.Reader, start time.Time, lineConv func([]byte) []byte, sink *usageSink) {
w := c.Writer
flusher, ok := w.(http.Flusher)
if !ok {
flusher = nopFlusher{}
}
// 原始响应捕获:仅管理员且开关开启(raw_request 已 set)时累积上游原始行
_, capture := c.Get("raw_request")
var rawResp strings.Builder
// commitRaw 在记账前把已累积的原始响应写入 context
commitRaw := func() {
if capture {
c.Set("raw_response", rawResp.String())
}
}
scanner := newSSEScanner(r)
for {
line, err := scanner.Next()
if line != nil {
if capture {
rawResp.Write(line)
}
out := line
if lineConv != nil {
out = lineConv(line)
}
if out != nil {
if _, werr := w.Write(out); werr != nil {
// 客户端意外断开:按已生成部分收费(canceled)
commitRaw()
g.finishUsage(c, ch, start, store.UsageStatusCanceled, "client_disconnect")
return
}
flusher.Flush()
if sink != nil {
sink.outputText += sseContentText(out)
}
}
if usageRaw := scanUsage(line); usageRaw != nil && sink != nil {
sink.push(usageRaw)
}
}
if err != nil {
commitRaw()
if err == io.EOF {
g.finishUsage(c, ch, start, store.UsageStatusSuccess, "")
} else if c.Request.Context().Err() != nil {
// 客户端意外断开:按已生成部分收费(canceled)
g.finishUsage(c, ch, start, store.UsageStatusCanceled, "client_disconnect")
} else {
g.recordError(c, ch, nil, start, "stream_read_error")
}
return
}
}
}
type nopFlusher struct{}
func (nopFlusher) Flush() {}
// ---------------------------------------------------------------------------
// usage 提取
// usageShape 兼容 chat (prompt/completion)、responses (input/output)、messages (input/output) 命名。
type usageShape struct {
PromptTokens int64 `json:"prompt_tokens"`
CompletionTokens int64 `json:"completion_tokens"`
InputTokens int64 `json:"input_tokens"`
OutputTokens int64 `json:"output_tokens"`
TotalTokens int64 `json:"total_tokens"`
CacheReadInputTokens int64 `json:"cache_read_input_tokens"`
CacheCreationInputTokens int64 `json:"cache_creation_input_tokens"`
}
// extractUsage 从完整响应体提取 usage 子对象(chat / responses / messages)。
func extractUsage(data []byte) json.RawMessage {
var m map[string]json.RawMessage
if json.Unmarshal(data, &m) != nil {
return nil
}
if u := usageFromMap(m); u != nil {
return u
}
if respRaw, ok := m["response"]; ok {
var resp map[string]json.RawMessage
if json.Unmarshal(respRaw, &resp) == nil {
if u := usageFromMap(resp); u != nil {
return u
}
}
}
if choices, ok := m["choices"]; ok {
var cs []map[string]json.RawMessage
if json.Unmarshal(choices, &cs) == nil {
for _, ch := range cs {
if msgRaw, ok := ch["message"]; ok {
var msg map[string]json.RawMessage
if json.Unmarshal(msgRaw, &msg) == nil {
if u := usageFromMap(msg); u != nil {
return u
}
}
}
}
}
}
return nil
}
// scanUsage 从 SSE 一行中提取 usage(OpenAI 末块 / responses completed / messages message_delta 等)。
func scanUsage(line []byte) json.RawMessage {
if !bytes.Contains(line, []byte(`"usage"`)) {
return nil
}
s, ok := sseDataPayload(line)
if !ok {
return nil
}
var m map[string]json.RawMessage
if json.Unmarshal([]byte(s), &m) != nil {
return nil
}
if u := usageFromMap(m); u != nil {
return u
}
if respRaw, ok := m["response"]; ok {
var resp map[string]json.RawMessage
if json.Unmarshal(respRaw, &resp) == nil {
if u := usageFromMap(resp); u != nil {
return u
}
}
}
return nil
}
// usageFromMap 从 map 顶层、message 或 delta 子对象中取 usage。
// 覆盖三种形态:
// - OpenAI 流式末块顶层 usage
// - Anthropic 非流式 / message_start 的 message.usage
// - Anthropic 流式 message_delta 的 delta.usage(真实 token 计数所在)
func usageFromMap(m map[string]json.RawMessage) json.RawMessage {
if u, ok := m["usage"]; ok && string(u) != "null" {
return u
}
if msgRaw, ok := m["message"]; ok {
var msg map[string]json.RawMessage
if json.Unmarshal(msgRaw, &msg) == nil {
if u, ok := msg["usage"]; ok && string(u) != "null" {
return u
}
}
}
if deltaRaw, ok := m["delta"]; ok {
var delta map[string]json.RawMessage
if json.Unmarshal(deltaRaw, &delta) == nil {
if u, ok := delta["usage"]; ok && string(u) != "null" {
return u
}
}
}
return nil
}
// sseScanner 按 SSE 行边界读取(兼容 \n 与 \r\n),保留原始行内容。
type sseScanner struct {
r *bufio.Reader
}
func newSSEScanner(r io.Reader) *sseScanner { return &sseScanner{r: bufio.NewReaderSize(r, 32*1024)} }
func (s *sseScanner) Next() ([]byte, error) {
line, err := s.r.ReadBytes('\n')
if len(line) > 0 {
return line, nil
}
if err != nil {
return nil, err
}
return nil, io.EOF
}
// ---------------------------------------------------------------------------
// 记账
// usageSink 累积多次 usage:合并各事件字段(message_start 给 input,message_delta 给 output)。
// outputText 累积已转发内容文本,用于流式中断时按 tiktoken 估算输出 token。
type usageSink struct {
us usageShape
outputText string
}
func (u *usageSink) push(raw json.RawMessage) {
if len(raw) == 0 {
return
}
var t usageShape
if json.Unmarshal(raw, &t) != nil {
return
}
// 零值不覆盖:不同事件携带不同字段
if t.PromptTokens > 0 {
u.us.PromptTokens = t.PromptTokens
}
if t.CompletionTokens > 0 {
u.us.CompletionTokens = t.CompletionTokens
}
if t.InputTokens > 0 {
u.us.InputTokens = t.InputTokens
}
if t.OutputTokens > 0 {
u.us.OutputTokens = t.OutputTokens
}
if t.CacheReadInputTokens > 0 {
u.us.CacheReadInputTokens = t.CacheReadInputTokens
}
if t.CacheCreationInputTokens > 0 {
u.us.CacheCreationInputTokens = t.CacheCreationInputTokens
}
}
// Shape 返回合并后的用量。
func (u *usageSink) Shape() usageShape { return u.us }
// finishUsage 落账:计算成本并异步写入。
func (g *Gateway) finishUsage(c *gin.Context, ch *store.Channel, start time.Time, status, errCode string) {
uid, _ := c.Get(CtxUserID)
kid, _ := c.Get(CtxKeyID)
trace, _ := c.Get(CtxTrace)
var us usageShape
var sink *usageSink
if h, ok := c.Get("usage_raw"); ok {
if holder, ok := h.(*sinkHolder); ok && holder.sink != nil {
us = holder.sink.Shape()
sink = holder.sink
}
}
in := us.PromptTokens + us.InputTokens
out := us.CompletionTokens + us.OutputTokens
cacheRead := us.CacheReadInputTokens
cacheCreate := us.CacheCreationInputTokens
modelName, _ := c.Get("model_name")
mn, _ := modelName.(string)
// 上游未返回真实 usage 时估算(tiktoken):
// - 流式中断(canceled):上游最终 usage 可能未返回
// - 流式正常结束(success):部分上游(如火山方舟 Anthropic 流式)message_start.usage 恒为 0、
// message_delta 不带 usage,只能按已收发内容估算,否则记账为 0 消耗
// 非流式上游必返回 usage,此处 in/out 非 0 不受影响。
if status == store.UsageStatusCanceled || status == store.UsageStatusSuccess {
if in == 0 {
if est, ok := c.Get("est_input_text"); ok {
if v, ok2 := est.(string); ok2 && v != "" {
in = int64(tokenizer.Count(v, mn))
}
}
}
if out == 0 && sink != nil && sink.outputText != "" {
out = int64(tokenizer.Count(sink.outputText, mn))
}
}
var model store.Model
var cost float64
var modelID uint64
_ = g.db.Where("name = ?", mn).First(&model).Error
if model.ID > 0 {
modelID = model.ID
cost = float64(in)/1e6*model.InputPrice +
float64(out)/1e6*model.OutputPrice +
float64(cacheRead)/1e6*model.CacheReadPrice
} else {
cost = float64(in)/1e6*0.15 + float64(out)/1e6*0.60 // 无定价模型时按示例价
}
proto, _ := c.Get("protocol")
p, _ := proto.(string)
if p == "" {
p = "chat"
}
traceStr, _ := trace.(string)
latency := int(time.Since(start).Milliseconds())
if status == store.UsageStatusSuccess && c.Writer.Status() >= 400 {
status = store.UsageStatusError
}
var errCodePtr *string
if errCode != "" {
errCodePtr = &errCode
}
var uidVal, kidVal uint64
if u, ok := uid.(uint64); ok {
uidVal = u
}
if k, ok := kid.(uint64); ok {
kidVal = k
}
var chID uint64
if ch != nil {
chID = ch.ID
}
var rawReq, rawResp string
if v, ok := c.Get("raw_request"); ok {
rawReq, _ = v.(string)
}
if v, ok := c.Get("raw_response"); ok {
rawResp, _ = v.(string)
}
// 密钥今日 token 用量累计(配额检查用)
if g.lim != nil && kidVal > 0 {
g.lim.AddTokens(kidVal, in+out)
}
g.rec.Record(&store.UsageLog{
RequestID: fmt.Sprintf("trace-%s", traceStr),
TraceID: traceStr,
UserID: uidVal,
KeyID: kidVal,
ChannelID: chID,
ModelID: modelID,
ModelName: mn,
Protocol: p,
InputTokens: in,
OutputTokens: out,
CacheReadTokens: cacheRead,
CacheCreationTokens: cacheCreate,
InputPrice: model.InputPrice,
OutputPrice: model.OutputPrice,
CacheReadPrice: model.CacheReadPrice,
Cost: cost,
LatencyMS: latency,
Status: status,
ErrorCode: errCodePtr,
RawRequest: rawReq,
RawResponse: rawResp,
CreatedAt: time.Now().UTC(),
})
}
// recordError 失败请求的记账(不产生扣费,status=error)。
func (g *Gateway) recordError(c *gin.Context, ch *store.Channel, resp *http.Response, start time.Time, code string) {
_ = resp
g.finishUsage(c, ch, start, store.UsageStatusError, code)
}
func now() time.Time { return time.Now() }